Google turns a video edit into a follow-up call against the previous interaction id

The load-bearing detail in the documentation is a parameter name. In the scene-extension sample, an edit is a fresh call carrying `previous_interaction_id` and a line of text such as "Continue the scene." ^[7] The asset is not re-uploaded; the server already holds it, addressed by the id of the interaction that produced it ^[7]. That is what makes conversational editing cheap to express ^[3], and it is also where the state now lives.

The context number is the substantive upgrade. Omni 1.1 reads up to ten seconds of preceding footage before continuing a shot, where Google says earlier models referenced only the final second ^[8], giving the model ten times the temporal run-up when it decides what happens next ^[9]. Consistency across a join is largely a function of how much of the past the generator can see, so this is the change most likely to show up in your output rather than in a slide.

Then read the ceiling. Extensions come in ten-second increments up to a cumulative forty seconds ^[10], which allows at most three extensions past an opening ten-second clip ^[11]. The unit of output is a shot or a very short sequence. A three-minute explainer still needs a timeline and someone to cut it; what the API now handles is the render, while the edit decision list still sits with you.

For anyone budgeting a creative tool, the cost line matters more than the 4K upscale ^[17]. Google puts 360p drafts at a third of the price of 720p ^[12], which buys three exploratory rounds for the money one polished round used to cost ^[14]. Iteration count is what makes a creative tool feel usable, and that is the parameter governing it. The speed half of the same sentence needs its footnote: "up to 60% faster" is stated on system throughput of 360p against 720p ^[13]. Throughput is a fleet measure. A pipeline clearing more jobs per second does not promise that your one request returns 60% sooner, and no median or tail latency is published to check it against.

That is the general shape of what is missing. Google calls the release production-ready for professional use through the Gemini API in AI Studio ^[20]. That claim describes intent; it is not a measurement. Across the launch post and the docs there is no eval, no human preference study, no comparison on shared prompts, and no absolute price or latency ^[21]. For a model whose pitch rests on physical plausibility and cultural context ^[1][22], the denominator is the thing to ask for: how many generations, judged by how many people, against what alternative. Absent that, the evaluation falls to you, and it should run on your own footage and your own prompts, not on the sample about iridescent marine diatoms ^[23].

One integration note before scoping the work. Direct REST callers receive the video as base64 inside a `steps` array, alongside a `thought` step, and the SDK's `output_video` shortcut does not exist at that layer ^[5][6]. The bytes arrive in the response body.

My read, with its condition attached: if your product is short-form and keeps a human reviewing drafts, this collapses a pipeline you were maintaining yourself. If it is long-form, the forty-second cumulative cap ^[10] means you are buying a shot generator and keeping the editor.